I Taught My Agent to Do the Part of My Job I Kept Skipping

I haven't manually updated a Salesforce record in weeks.

It started with a call that seemed like nothing special. Their CTO asked three sharp questions about SSO, the budget number had gone up since we last talked, and the decision had slipped from this quarter into next. Granola had all of it in a clean summary the moment the call ended. Salesforce still showed what I'd left it with after the last call, because I hadn't touched it since.

I used to tell myself I'd go fix it before the next meeting. I never did. By the time I opened Salesforce, usually days later, I was typing from memory instead of the notes. And I wasn't the only one doing this. Everyone kept the real version in their head, and Salesforce got whatever they remembered, whenever they got around to it. So the pipeline review was part real, part memory. Nobody planned it that way. It happened on its own, one missed update at a time.

So that Friday I stopped doing it myself and gave my agent the two tools instead, Granola through its API, Salesforce through an OAuth connection in OneCLI. The next call went the same way it always had, except this time I didn't have to do anything once it ended. The summary landed, and the agent logged the call and updated the deal to match, exactly the way I would have if I'd sat down and done it myself.

One morning I noticed a task it had opened the day before: due Friday, tied to a security doc. It landed in Salesforce, though it could just as easily have gone to Jira or Monday. I went back and checked the call. Someone had mentioned it once, near the end, almost as an afterthought, exactly the kind of line I would have missed. The agent hadn't.

It gets better at being me

The first few calls, the agent did the work the way a new hire does the work: correctly, but literally. It logged what was said and updated what was explicit, because that's all it knew about us.

Then it started to know more. It has every call I've been on, every summary, every fix I made to its work. It picked up how we run deals, what usually happens after a first call with a technical buyer, which follow-ups we always send, how our team talks about pricing. The fixes I made to its updates went from rewrites, to tweaks, to mostly nothing.

That's the same curve an experienced employee rides, and I treat it the same way. The routine work runs on its own now, logging calls, updating next steps, opening the follow-ups, because it keeps getting it right. The point is that I didn't decide to trust it. It showed me I could.

It's never seen my Salesforce password either. The connection lives in OneCLI, the agent gets access to it, and every action it takes is logged.

Now the call ends, I close the laptop, and for once that's actually the end of it.